This site contains C++ codes for training nonlinear SVMs via stochastic approximation methods described in

• Sangkyun Lee, and Stephen J. Wright, " ASSET: Approximate Stochastic Subgradient Estimation Training for Support Vector Machines ", ICPRAM, 2012.

Software

• Asset v0.1 [ download ]
• Asset is based on the following softwares:
  • Pegasos SVM solver v1.0 [link]
  • LAPACK [link]
  • agner.org random number generator [link]

• Usage
  • asset [options] [-modelFile model_file] [-testFile test_file] train_file
  • options:

    -algo [number]
    	4: Batch mode, strongly convex
    	5: Batch mode, general convex
    	6: Online mode, strongly convex
    	7: Online mode, general convex
    -b	:Use bias term or not {0,1} (default: 0)
    -lambda	:Regularization parameter (default: 0.001)
    -iter	: Number of maximal iterations to run (default: 1/lambda)
    -g	: RBF kernel bandwidth exp(-g||x-y||^2) (default: 1.0)
    -rank	: Rank of kernel approximation (default: 512)
    -sol_eps	: Kernel expansion coefficient threshold (default: 1e-12)
    -target_err	: Target test error to stop algorithms

License

• This software is open-source under the agreement of GPL v2.0.